[compiler-rt] replicate changes from llvm/ProfileData/InstrProfData.inc

[andrurogerz]

Purpose

The compiler-rt project check-same-common-code.test test case started failing after #142861 was merged. This change addresses the failure.

Overview

This patch replicates the changes made by #142861 to llvm/include/llvm/ProfileData/InstrProfData.inc in the duplicated file compiler-rt/include/profile/InstrProfData.inc. These files otherwise match.

Validation

Locally built check-profile target and verified check-same-common-code.test now passes.

[compnerd]

I think that we can merge this for now, but we will need to revisit this. I'm not sure that we want to have this copy dllimport symbols from within LLVM.

[boomanaiden154]

Where exactly does LLVM_ABI get defined in compiler-rt?

[andrurogerz]

Where exactly does LLVM_ABI get defined in compiler-rt?

It doesn't; it is relying on picking it up from llvm/include/Support/Compiler.h. @boomanaiden154 if you don't think this is a good idea, or is incorrect, I am happy to revert this change along with the change to the file it is duped from, llvm\ProfileData\InstrProfData.inc.

[compnerd]

@boomanaiden154 this is basically just to appease the diff check; the dll annotations aren't being applied to the profile runtime. I don't think that the profile runtime should be exposing these interfaces ever (the runtime's ABI is a pure C interface).

[boomanaiden154]

It doesn't; it is relying on picking it up from llvm/include/Support/Compiler.h

And that's getting pulled into the compiler-rt build tree? I'm mostly just wondering why we aren't running into build errors when compiling compiler-rt.

this is basically just to appease the diff check; the dll annotations aren't being applied to the profile runtime. I don't think that the profile runtime should be exposing these interfaces ever (the runtime's ABI is a pure C interface).

Yeah, I got that this was to appease the diff check. I agree it doesn't make sense to export them from the runtime given it's always (?) statically linked and a pure C interface. Having LLVM_ABI be noop on the compiler-rt side for the common headers seems reasonable enough to me (or some other annotation in the file that maps to LLVM_ABI in certain contexts that is more clear). I guess stripping LLVM_ABI and then running diff would work too.